7 Refund requests. 7 taps on Approve. $214 of it shouldn't have gone out.
"Refund request: Order #1182, $64. Approve?"
That card hit the phone of Mara, owner of Quillmoss Candles (a made-up online candle shop), on Monday. She tapped Approve. On Tuesday the same card came back from a chat ticket. She tapped Approve again. Same customer, same order, $64 x 2 = $128 out the door.
By Sunday she had approved 7 refunds worth $404. Only $190 should have gone out.
An Approve button is only as safe as what's printed on it.
AI Prompts, Reviewed takes one prompt or workflow from How I AI, Claire Vo's show where guests share their screens and walk through how they really use AI. We run it on a small-business job, score it, and tell you what we'd change.
Credit: Based on "How to Automate Customer Support with a 5-Star AI Agent," a workflow from How I AI (hosted by Claire Vo, written up by the ChatPRD team), shown by Claire Vo herself in the episode "Practical Grok Bot Agents and Workflows That Replaced OpenClaw." Original workflow: https://www.chatprd.ai/how-i-ai/workflows/how-to-automate-customer-support-with-a-5-star-ai-agent. Full episode: https://www.chatprd.ai/how-i-ai/grok-bot-agent-templates-and-workflows. Guest site: none listed on the How I AI workflow page.
What the original does
Claire builds a support bot in Grok Bot, names it something like "Holly Helpdesk," and connects it to Intercom, email, and Slack. It sweeps the inbox on a schedule (hourly is the example), answers from your support playbook and docs, and once a week suggests new help articles. The smart part is money: the bot can spot a refund request but should "not issue them unilaterally." It sends you an Approve button tied to a single-use action in Stripe.
The test
Everything here is made up. Quillmoss sells candles online, about 40 orders a week. Policy: refunds within 30 days of delivery. One week, Monday, Oct 12 to Sunday, Oct 18, 2026.
Naive run (inbox and chat connected, "send me refund requests to approve"). Each card showed an order number, an amount, and a one-line reason. Mara approved all 7. Three were wrong:
- Paid twice. Order #1182 came in by email Monday and by chat Tuesday. Two tickets, two cards, $64 refunded twice.
- Outside policy. Order #1097, $112, delivered Friday, Aug 28, request Wednesday, Oct 14. That's 47 days, not 30. The card said "arrived damaged," with no dates.
- Wrong fix. Order #1210, $38, cracked jar. The customer asked for a replacement. The bot filed it as a refund. Now Mara owes a new candle too.
Wrong money: $64 + $112 + $38 = $214. If that's a normal week, $214 x 52 = $11,128 a year.
The rerun
Same week, plus 3 rules: one card per order, never per ticket. Every card shows delivery date, days since delivery against the policy, past refunds on that order, and what the customer actually asked for. The bot recommends APPROVE, DECLINE, or ASK, with the reason.
Result: 5 refund cards and 1 replacement note. Mara approved 4 for $42 + $29 + $55 + $64 = $190. #1097 came up DECLINE: 47 days against a 30-day policy. She offered store credit.
Verdict: 7/10
Strengths: the human approves the money, and the single-use Stripe action limits the damage. The weekly "what should we add to the docs" routine cuts repeat tickets.
Weakness: the page never says what the Approve card should show, how to catch two tickets about one order, or how to check your refund policy. Thin cards train you to tap fast.
Founder tweak: group by order, put the dates and the policy math on the card, and make "replacement" and "store credit" their own buttons.
Copy the MCAFax version
This is our own rewrite, not the original. Copy the original from Claire Vo on How I AI: https://www.chatprd.ai/how-i-ai/workflows/how-to-automate-customer-support-with-a-5-star-ai-agent
You are the support assistant for [business name].
You draft replies and refund requests. You never send money.
Inputs I will connect or paste:
(A) support inbox, chat, and order history
(B) my refund policy: [days] days from delivery, [conditions]
(C) my support playbook and help docs
Step 1. Group every new ticket by order number.
Two tickets on one order = one case. Say so.
Step 2. For each case, decide what the customer asked for:
REFUND, REPLACEMENT, STORE CREDIT, or QUESTION.
Quote their words. Do not upgrade a question to a refund.
Step 3. For refunds, build one approval card:
Order #, amount, order date, delivery date,
Days since delivery vs policy (show the math),
Refunds already paid on this order,
Recommendation: APPROVE, DECLINE, or ASK, with one reason.
Step 4. For everything else, draft a reply from (C).
If (C) doesn't cover it, say "not in playbook."
Every Friday: list the 3 questions asked most this week
That (C) doesn't answer, with ticket counts.
Paste into Claude, ChatGPT, or Gemini with your inbox and orders connected or exported.
Before you trust it
- Write your refund policy in days, not "a reasonable time."
- Check that every card shows the delivery date. If it doesn't, don't tap.
- Search for a second ticket on the same order before approving.
- Give replacement and store credit their own path.
- Turn the Friday list into help articles: https://mcafax.com/notebook/prompts/faq-help-center-builder
Want a support chat on your site too? https://mcafax.com/notebook/prompts/customer-support-chat-widget
More owner prompts: https://mcafax.com/notebook/prompts
FAQ
What do I need for the original workflow? A Grok Bot agent connected to your support tools (Intercom, email, Slack), your support playbook and docs, and Stripe for the refund approval step.
Can the AI issue refunds on its own? Not in the original. It finds refund requests and sends you an Approve button. Keep it that way.
Why did one customer get refunded twice? They wrote in by email and by chat, and the bot made one card per ticket. Group by order number.
What should a refund approval card show? Order number, amount, delivery date, days since delivery against your policy, refunds already paid, and what the customer asked for.
How often should the bot check the inbox? The page's example is hourly. Pick a pace you'll actually answer, since every card waits on you.
Can it cut my ticket volume? That's the weekly step: it lists questions your docs don't answer, so you can add them.